perf: Support a new coresum event qualifier
cover perf: Support a new coresum event qualifier
The coresum event qualifier which sums up the event counts for both hardware threads in a core. For example,
perf stat -e cpu/event=0,umask=0x3,coresum=1/,cpu/event=0,umask=0x3/
In this example, we count the event 'ref-cycles' per-core and per-CPU in one perf stat command-line.
We can already support per-core counting with --per-core, but it's often useful to do this together with other metrics that are collected per CPU (per hardware thread). So this patch series supports this per-core counting on a event level.
Jin Yao (3):
perf: Add a coresum event qualifier
perf stat: Support coresum event qualifier
perf test: Add a simple test for term coresum
tools/perf/Documentation/perf-stat.txt | 4 ++ tools/perf/builtin-stat.c | 21 +++++++ tools/perf/tests/parse-events.c | 10 ++- tools/perf/util/evsel.c | 2 + tools/perf/util/evsel.h | 3 + tools/perf/util/parse-events.c | 27 +++++++++ tools/perf/util/parse-events.h | 1 + tools/perf/util/parse-events.l | 1 + tools/perf/util/stat-display.c | 108 ++++++++++++++++++++++++--------- tools/perf/util/stat.c | 8 ++- 10 files changed, 151 insertions(+), 34 deletions(-)
Signature · 1 line
2.7.4
PATCH 1/3 perf: Add a coresum event qualifier
Add a coresum event qualifier, like cpu/event=0,umask=0x3,coresum=1/, that sums up the event counts for both hardware threads in a core.
We can already do this with --per-core, but it's often useful to do this together with other metrics that are collected per hardware thread. So we need to support this per-core counting on a event level.
This can be implemented in only the user tool, no kernel support needed.
--- tools/perf/util/evsel.c | 2 ++ tools/perf/util/evsel.h | 3 +++ tools/perf/util/parse-events.c | 27 +++++++++++++++++++++++++++ tools/perf/util/parse-events.h | 1 + tools/perf/util/parse-events.l | 1 + 5 files changed, 34 insertions(+)
tools/perf/util/evsel.c +2 -0
@@ -803,6 +803,8 @@ static void apply_config_terms(struct perf_evsel *evsel,break ;case PERF_EVSEL__CONFIG_TERM_DRV_CFG :break ; +case PERF_EVSEL__CONFIG_TERM_CORESUM : +break ;default :break ; }
tools/perf/util/evsel.h +3 -0
@@ -50,6 +50,7 @@ enum term_type {PERF_EVSEL__CONFIG_TERM_OVERWRITE ,PERF_EVSEL__CONFIG_TERM_DRV_CFG ,PERF_EVSEL__CONFIG_TERM_BRANCH , + PERF_EVSEL__CONFIG_TERM_CORESUM, }; struct perf_evsel_config_term {
@@ -67,6 +68,7 @@ struct perf_evsel_config_term {bool overwrite;char *branch;unsigned long max_events; +bool coresum; } val;bool weak; };
@@ -150,6 +152,7 @@ struct perf_evsel {struct perf_evsel **metric_events;bool collect_stat;bool weak_group; +bool coresum;const char *pmu_name; };
tools/perf/util/parse-events.c +27 -0
@@ -930,6 +930,7 @@ static const char *config_term_names[__PARSE_EVENTS__TERM_TYPE_NR] = { [PARSE_EVENTS__TERM_TYPE_OVERWRITE] = "overwrite", [PARSE_EVENTS__TERM_TYPE_NOOVERWRITE] = "no-overwrite", [PARSE_EVENTS__TERM_TYPE_DRV_CFG] = "driver-config", + [PARSE_EVENTS__TERM_TYPE_CORESUM] = "coresum", };static bool config_term_shrinked;
@@ -950,6 +951,7 @@ config_term_avail(int term_type, struct parse_events_error *err)case PARSE_EVENTS__TERM_TYPE_CONFIG2 :case PARSE_EVENTS__TERM_TYPE_NAME :case PARSE_EVENTS__TERM_TYPE_SAMPLE_PERIOD : +case PARSE_EVENTS__TERM_TYPE_CORESUM :return true;default :if (!err)
@@ -1041,6 +1043,14 @@ do { \case PARSE_EVENTS__TERM_TYPE_MAX_EVENTS :CHECK_TYPE_VAL (NUM );break ; +case PARSE_EVENTS__TERM_TYPE_CORESUM : +CHECK_TYPE_VAL (NUM ); +if ((unsigned int )term->val.num >1 ) { + err->str =strdup ("expected 0 or 1" ); + err->idx = term->err_val; +return -EINVAL ; + } +break ;default : err->str =strdup ("unknown term" ); err->idx = term->err_term;
@@ -1179,6 +1189,10 @@ do { \case PARSE_EVENTS__TERM_TYPE_DRV_CFG :ADD_CONFIG_TERM (DRV_CFG , drv_cfg, term->val.str);break ; +case PARSE_EVENTS__TERM_TYPE_CORESUM : +ADD_CONFIG_TERM (CORESUM , coresum, + term->val.num ?1 :0 ); +break ;default :break ; }
@@ -1233,6 +1247,18 @@ int parse_events_add_numeric(struct parse_events_state *parse_state, get_config_name(head_config), &config_terms); } +static bool config_term_coresum (struct list_head *config_terms) +{ +struct perf_evsel_config_term *term; + +list_for_each_entry (term, config_terms, list) { +if (term->type ==PERF_EVSEL__CONFIG_TERM_CORESUM ) +return term->val.coresum ? true : false; + } + +return false; +} + int parse_events_add_pmu(struct parse_events_state *parse_state, struct list_head *list, char *name, struct list_head *head_config,
@@ -1305,6 +1331,7 @@ int parse_events_add_pmu(struct parse_events_state *parse_state, evsel->metric_name = info.metric_name; evsel->pmu_name = name; evsel->use_uncore_alias = use_uncore_alias; + evsel->coresum =config_term_coresum (&evsel->config_terms); }return evsel ?0 : -ENOMEM ;
tools/perf/util/parse-events.h +1 -0
@@ -75,6 +75,7 @@ enum {PARSE_EVENTS__TERM_TYPE_NOOVERWRITE ,PARSE_EVENTS__TERM_TYPE_OVERWRITE ,PARSE_EVENTS__TERM_TYPE_DRV_CFG , +PARSE_EVENTS__TERM_TYPE_CORESUM , __PARSE_EVENTS__TERM_TYPE_NR, };
tools/perf/util/parse-events.l +1 -0
@@ -274,6 +274,7 @@ inherit { return term(yyscanner, PARSE_EVENTS__TERM_TYPE_INHERIT); } no-inherit { return term(yyscanner, PARSE_EVENTS__TERM_TYPE_NOINHERIT); } overwrite { return term(yyscanner, PARSE_EVENTS__TERM_TYPE_OVERWRITE); } no-overwrite { return term(yyscanner, PARSE_EVENTS__TERM_TYPE_NOOVERWRITE); } +coresum { return term(yyscanner, PARSE_EVENTS__TERM_TYPE_CORESUM); } , { return ','; } "/" { BEGIN(INITIAL); return '/'; } {name_minus} { return str(yyscanner, PE_NAME); }
Signature · 1 line
2.7.4
PATCH 2/3 perf stat: Support coresum event qualifier
In reply to Jin YaoWith this patch, we can use the coresum event qualifier in perf-stat.
root@skl:/tmp# perf stat -e cpu/event=0,umask=0x3,coresum=1/,cpu/event=0,umask=0x3/ -a -A -I1000
# time CPU counts unit events
1.000816660 S0-C0 80,676,154 cpu/event=0,umask=0x3,coresum=1/ (50.04%)
1.000816660 S0-C1 115,314,959 cpu/event=0,umask=0x3,coresum=1/ (50.04%)
1.000816660 S0-C2 126,541,249 cpu/event=0,umask=0x3,coresum=1/ (50.04%)
1.000816660 S0-C3 119,950,015 cpu/event=0,umask=0x3,coresum=1/ (50.04%)
1.000816660 CPU0 52,439,489 cpu/event=0,umask=0x3/ (49.96%)
1.000816660 CPU1 53,431,155 cpu/event=0,umask=0x3/ (49.96%)
1.000816660 CPU2 91,192,070 cpu/event=0,umask=0x3/ (49.96%)
1.000816660 CPU3 90,852,159 cpu/event=0,umask=0x3/ (49.96%)
1.000816660 CPU4 29,715,956 cpu/event=0,umask=0x3/ (49.96%)
1.000816660 CPU5 63,289,507 cpu/event=0,umask=0x3/ (49.96%)
1.000816660 CPU6 29,036,120 cpu/event=0,umask=0x3/ (49.96%)
1.000816660 CPU7 28,996,591 cpu/event=0,umask=0x3/ (49.97%)In this example, we count the event 'ref-cycles' per-core and per-CPU in one perf stat command-line. From the output, we can see:
S0-C0 = CPU0 + CPU4
S0-C1 = CPU1 + CPU5
S0-C2 = CPU2 + CPU6
S0-C3 = CPU3 + CPU7
So the result is expected (tiny difference is ignored).
--- tools/perf/Documentation/perf-stat.txt | 4 ++ tools/perf/builtin-stat.c | 21 +++++++ tools/perf/util/stat-display.c | 108 ++++++++++++++++++++++++--------- tools/perf/util/stat.c | 8 ++- 4 files changed, 108 insertions(+), 33 deletions(-)
tools/perf/Documentation/perf-stat.txt +4 -0
@@ -43,6 +43,10 @@ report:: param1 and param2 are defined as formats for the PMU in /sys/bus/event_source/devices/<pmu>/format/* + 'coresum' is a event qualifier that sums up the event counts for both + hardware threads in a core. For example: + perf stat -A -a -e cpu/event,coresum=1/,otherevent ... + - a symbolically formed event like 'pmu/config=M,config1=N,config2=K/' where M, N, K are numbers (in decimal, hex, octal format). Acceptable values for each of 'config', 'config1' and 'config2'
tools/perf/builtin-stat.c +21 -0
@@ -832,6 +832,18 @@ static int perf_stat__get_core_cached(struct perf_stat_config *config,return perf_stat__get_aggr (config, perf_stat__get_core, map, idx); } +static bool term_coresum_set (void ) +{ +struct perf_evsel *counter; + +evlist__for_each_entry (evsel_list, counter) { +if (counter->coresum) +return true; + } + +return false; +} +static int perf_stat_init_aggr_mode (void ) {int nr;
@@ -852,6 +864,15 @@ static int perf_stat_init_aggr_mode(void) stat_config.aggr_get_id = perf_stat__get_core_cached;break ;case AGGR_NONE : +if (term_coresum_set ()) { +if (cpu_map__build_core_map (evsel_list->cpus, + &stat_config.aggr_map)) { +perror ("cannot build core map" ); +return -1 ; + } + stat_config.aggr_get_id = perf_stat__get_core_cached; + } +break ;case AGGR_GLOBAL :case AGGR_THREAD :case AGGR_UNSET :
tools/perf/util/stat-display.c +78 -30
151 lines not on this page · raw message
tools/perf/util/stat.c +5 -3
@@ -277,9 +277,11 @@ process_counter_values(struct perf_stat_config *config, struct perf_evsel *evselif (!evsel->snapshot)perf_evsel__compute_deltas (evsel, cpu, thread, count);perf_counts_values__scale (count, config->scale,NULL ); -if (config->aggr_mode ==AGGR_NONE ) -perf_stat__update_shadow_stats (evsel, count->val, cpu, - &rt_stat); +if ((config->aggr_mode ==AGGR_NONE ) && (!evsel->coresum)) { +perf_stat__update_shadow_stats (evsel, count->val, + cpu, &rt_stat); + } +if (config->aggr_mode ==AGGR_THREAD ) {if (config->stats) perf_stat__update_shadow_stats(evsel,
Signature · 1 line
2.7.4
Re: [PATCH v1 2/3] perf stat: Support coresum event qualifier
On Sat, Mar 16, 2019 at 12:04:15AM +0800, Jin Yao wrote:
SNIP
Quoted 24 lines
tools/perf/util/stat-display.c +24
-0+static void print_counter_aggrdata (struct perf_stat_config *config, +struct perf_evsel *counter,int s, +char *prefix,bool metric_only, +bool *first) +{ +struct aggr_data ad; +FILE *output = config->output; +u64 ena, run, val; +int id, nr; +double uval; + + ad.id = id = config->aggr_map->map[s]; + ad.val = ad.ena = ad.run =0 ; + ad.nr =0 ; +if (!collect_data (config, counter, aggr_cb, &ad)) +return ; + + nr = ad.nr; + ena = ad.ena; + run = ad.run; + val = ad.val; +if (*first && metric_only) { + *first = false; +aggr_printout (config, counter, id, nr);
tools/perf/util/stat-display.c static void aggr_cb(struct perf_stat_config *config, · line 643
Quoted 10 lines
tools/perf/util/stat-display.c +10
-0+ } +if (prefix && !metric_only) +fprintf (output,"%s" , prefix); + + uval = val * counter->scale; +printout (config, id, nr, counter, uval, prefix, + run, ena,1.0 , &rt_stat); +if (!metric_only) +fputc ('\n' , output); +}
plese put the factoring out of print_counter_aggrdata function into separate patch
thanks,
jirka
Quoted 52 lines
tools/perf/util/stat-display.c +5
-26+static void print_aggr (struct perf_stat_config *config,struct perf_evlist *evlist,char *prefix)@@ -606,9 +649,7 @@ static void print_aggr(struct perf_stat_config *config,bool metric_only = config->metric_only;FILE *output = config->output;struct perf_evsel *counter; -int s, id, nr; -double uval; -u64 ena, run, val; +int s;bool first;if (!(config->aggr_map || config->aggr_get_id))@@ -621,36 +662,16 @@ static void print_aggr(struct perf_stat_config *config,* Without each counter has its own line. */ for (s =0 ; s < config->aggr_map->nr; s++) { -struct aggr_data ad;if (prefix && metric_only)fprintf (output,"%s" , prefix); - ad.id = id = config->aggr_map->map[s]; first = true;evlist__for_each_entry (evlist, counter) {if (is_duration_time (counter))continue ; - - ad.val = ad.ena = ad.run =0 ; - ad.nr =0 ; -if (!collect_data (config, counter, aggr_cb, &ad)) -continue ; - nr = ad.nr; - ena = ad.ena; - run = ad.run; - val = ad.val; -if (first && metric_only) { - first = false; -aggr_printout (config, counter, id, nr); - } -if (prefix && !metric_only) -fprintf (output,"%s" , prefix); - - uval = val * counter->scale; -printout (config, id, nr, counter, uval, prefix, - run, ena,1.0 , &rt_stat); -if (!metric_only) -fputc ('\n' , output); +print_counter_aggrdata (config, counter, s, + prefix, metric_only, + &first);
SNIP
Re: [PATCH v1 2/3] perf stat: Support coresum event qualifier
Quoted 42 lines · Jiri Olsa
On Sat, Mar 16, 2019 at 12:04:15AM +0800, Jin Yao wrote:
SNIP
Quoted 24 lines
tools/perf/util/stat-display.c +24
-0+static void print_counter_aggrdata (struct perf_stat_config *config, +struct perf_evsel *counter,int s, +char *prefix,bool metric_only, +bool *first) +{ +struct aggr_data ad; +FILE *output = config->output; +u64 ena, run, val; +int id, nr; +double uval; + + ad.id = id = config->aggr_map->map[s]; + ad.val = ad.ena = ad.run =0 ; + ad.nr =0 ; +if (!collect_data (config, counter, aggr_cb, &ad)) +return ; + + nr = ad.nr; + ena = ad.ena; + run = ad.run; + val = ad.val; +if (*first && metric_only) { + *first = false; +aggr_printout (config, counter, id, nr);tools/perf/util/stat-display.c static void aggr_cb(struct perf_stat_config *config, · line 643
Quoted 10 lines
tools/perf/util/stat-display.c +10
-0+ } +if (prefix && !metric_only) +fprintf (output,"%s" , prefix); + + uval = val * counter->scale; +printout (config, id, nr, counter, uval, prefix, + run, ena,1.0 , &rt_stat); +if (!metric_only) +fputc ('\n' , output); +}plese put the factoring out of
print_counter_aggrdatafunction into separate patch
Quoted 4 lines
thanks,
jirka
OK!
Thanks
Jin Yao
Quoted 55 lines
Quoted 52 lines
tools/perf/util/stat-display.c +5
-26+static void print_aggr (struct perf_stat_config *config,struct perf_evlist *evlist,char *prefix)@@ -606,9 +649,7 @@ static void print_aggr(struct perf_stat_config *config,bool metric_only = config->metric_only;FILE *output = config->output;struct perf_evsel *counter; -int s, id, nr; -double uval; -u64 ena, run, val; +int s;bool first;if (!(config->aggr_map || config->aggr_get_id))@@ -621,36 +662,16 @@ static void print_aggr(struct perf_stat_config *config,* Without each counter has its own line. */ for (s =0 ; s < config->aggr_map->nr; s++) { -struct aggr_data ad;if (prefix && metric_only)fprintf (output,"%s" , prefix); - ad.id = id = config->aggr_map->map[s]; first = true;evlist__for_each_entry (evlist, counter) {if (is_duration_time (counter))continue ; - - ad.val = ad.ena = ad.run =0 ; - ad.nr =0 ; -if (!collect_data (config, counter, aggr_cb, &ad)) -continue ; - nr = ad.nr; - ena = ad.ena; - run = ad.run; - val = ad.val; -if (first && metric_only) { - first = false; -aggr_printout (config, counter, id, nr); - } -if (prefix && !metric_only) -fprintf (output,"%s" , prefix); - - uval = val * counter->scale; -printout (config, id, nr, counter, uval, prefix, - run, ena,1.0 , &rt_stat); -if (!metric_only) -fputc ('\n' , output); +print_counter_aggrdata (config, counter, s, + prefix, metric_only, + &first);SNIP
PATCH 3/3 perf test: Add a simple test for term coresum
In reply to Jin YaoIt's a simple test which just checks if parser works.
--- tools/perf/tests/parse-events.c | 10 +++++++++- 1 file changed, 9 insertions(+), 1 deletion(-)
tools/perf/tests/parse-events.c +9 -1
@@ -605,6 +605,14 @@ static int test__checkterms_simple(struct list_head *terms)TEST_ASSERT_VAL ("wrong val" , term->val.num ==1 );TEST_ASSERT_VAL ("wrong config" , !strcmp (term->config,"umask" )); +/* coresum=1*/ + term = list_entry(term->list.next, struct parse_events_term, list); +TEST_ASSERT_VAL ("wrong type term" , + term->type_term ==PARSE_EVENTS__TERM_TYPE_CORESUM ); +TEST_ASSERT_VAL ("wrong type val" , + term->type_val ==PARSE_EVENTS__TERM_TYPE_NUM ); +TEST_ASSERT_VAL ("wrong val" , term->val.num ==1 ); +return 0 ; }
@@ -1734,7 +1742,7 @@ struct terms_test {static struct terms_test test__terms[] = { [0 ] = { - .str ="config=10,config1,config2=3,umask=1" , + .str ="config=10,config1,config2=3,umask=1,coresum=1" , .check = test__checkterms_simple, }, };
Signature · 1 line
2.7.4
Quoted 5 lines · Jin Yao
The coresum event qualifier which sums up the event counts for both hardware threads in a core. For example,
perf stat -e cpu/event=0,umask=0x3,coresum=1/,cpu/event=0,umask=0x3/
Quoted 7 lines
In this example, we count the event 'ref-cycles' per-core and per-CPU in one perf stat command-line.
We can already support per-core counting with --per-core, but it's often useful to do this together with other metrics that are collected per CPU (per hardware thread). So this patch series supports this per-core counting on a event level.
seems useful, but perhaps we should follow the --per-core option we already have and call it 'per-core' instead of coresum
jirka
Quoted 21 lines
Jin Yao (3):
perf: Add a coresum event qualifier
perf stat: Support coresum event qualifier
perf test: Add a simple test for term coresumtools/perf/Documentation/perf-stat.txt | 4 ++ tools/perf/builtin-stat.c | 21 +++++++ tools/perf/tests/parse-events.c | 10 ++- tools/perf/util/evsel.c | 2 + tools/perf/util/evsel.h | 3 + tools/perf/util/parse-events.c | 27 +++++++++ tools/perf/util/parse-events.h | 1 + tools/perf/util/parse-events.l | 1 + tools/perf/util/stat-display.c | 108 ++++++++++++++++++++++++--------- tools/perf/util/stat.c | 8 ++- 10 files changed, 151 insertions(+), 34 deletions(-)Signature · 1 line
2.7.4
Quoted 14 lines · Jiri Olsa
Quoted 5 lines · Jin Yao
The coresum event qualifier which sums up the event counts for both hardware threads in a core. For example,
perf stat -e cpu/event=0,umask=0x3,coresum=1/,cpu/event=0,umask=0x3/
Quoted 7 lines
In this example, we count the event 'ref-cycles' per-core and per-CPU in one perf stat command-line.
We can already support per-core counting with --per-core, but it's often useful to do this together with other metrics that are collected per CPU (per hardware thread). So this patch series supports this per-core counting on a event level.
Quoted 5 lines
seems useful, but perhaps we should follow the --per-core option we already have and call it 'per-core' instead of coresum
jirka
Yes, the coresum's behavior is similar as --per-core option, just supports at the event level. I'm OK with calling it 'per-core'.
For example,
perf stat -e cpu/event=0,umask=0x3,per-core=1/
Thanks
Jin Yao
Quoted 21 lines
Quoted 21 lines
Jin Yao (3):
perf: Add a coresum event qualifier
perf stat: Support coresum event qualifier
perf test: Add a simple test for term coresumtools/perf/Documentation/perf-stat.txt | 4 ++ tools/perf/builtin-stat.c | 21 +++++++ tools/perf/tests/parse-events.c | 10 ++- tools/perf/util/evsel.c | 2 + tools/perf/util/evsel.h | 3 + tools/perf/util/parse-events.c | 27 +++++++++ tools/perf/util/parse-events.h | 1 + tools/perf/util/parse-events.l | 1 + tools/perf/util/stat-display.c | 108 ++++++++++++++++++++++++--------- tools/perf/util/stat.c | 8 ++- 10 files changed, 151 insertions(+), 34 deletions(-)Signature · 1 line
2.7.4
Quoted 5 lines
Yes, the coresum's behavior is similar as --per-core option, just supports at the event level. I'm OK with calling it 'per-core'.
For example,
perf stat -e cpu/event=0,umask=0x3,per-core=1/
Please use percore, the - would need to be escaped in metric expressions.
-Andi
Quoted 10 lines · Andi Kleen
Quoted 5 lines
Yes, the coresum's behavior is similar as --per-core option, just supports at the event level. I'm OK with calling it 'per-core'.
For example,
perf stat -e cpu/event=0,umask=0x3,per-core=1/Please use percore, the - would need to be escaped in metric expressions.
-Andi
Oh, yes, thanks for reminding. Will use 'percore' in next version.
Thanks
Jin Yao